[Paper] A Two-Stage, Object-Centric Deep Learning Framework for Robust Exam Cheating Detection

Source: arXiv - 2604.16234v1

Overview

The paper presents a lightweight, two‑stage deep‑learning pipeline that can spot cheating behavior in exam rooms from video frames. By pairing a fast object detector (YOLOv8n) with a specialized behavior classifier (RexNet‑150), the authors achieve near‑human accuracy while keeping inference latency low enough for real‑time, large‑scale deployments.

Key Contributions

• Simple two‑stage architecture: Detects each student with YOLOv8n, then classifies the cropped region as “normal” or “cheating” using a fine‑tuned RexNet‑150 model.
• Large, heterogeneous dataset: 273 k+ samples gathered from 10 independent sources, providing diverse lighting, camera angles, and classroom layouts.
• High performance: 0.95 accuracy, 0.94 recall, 0.96 precision, and 0.95 F1‑score—about a 13 % boost over a strong video‑based baseline.
• Real‑time inference: Average processing time of 13.9 ms per frame, enabling live monitoring of thousands of seats simultaneously.
• Ethical delivery mechanism: Results are sent privately to each student (e.g., via email) to avoid public shaming and give individuals a chance to reflect.

Methodology

1. Stage 1 – Student Localization

   • Uses YOLOv8n (the “nano” version of YOLOv8) for its balance of speed and detection quality.
   • The model outputs bounding boxes around every person in an exam‑room image.

2. Stage 2 – Behavior Classification

   • Each bounding box is cropped, resized, and normalized.
   • A RexNet‑150 network—pre‑trained on ImageNet and then fine‑tuned on the cheating dataset—classifies the crop as normal or cheating.
   • The two stages are decoupled, so improvements to either detector or classifier can be swapped in without retraining the whole pipeline.

3. Training & Validation

   • The combined dataset is split into training/validation/test sets, preserving source diversity.
   • Standard augmentations (random flips, brightness jitter) help the classifier generalize to unseen camera conditions.

4. Deployment Considerations

   • The pipeline runs on commodity GPUs or even high‑end CPUs, thanks to the low‑parameter YOLOv8n and the efficient RexNet‑150 backbone.
   • Inference is performed per frame; temporal smoothing (e.g., majority vote over consecutive frames) is left as a future enhancement.

Results & Findings

• The system outperforms a strong video‑based baseline (0.82 accuracy) by 13 %, confirming that a well‑designed two‑stage image pipeline can rival more complex multi‑modal solutions.
• High precision means false accusations are rare, a crucial factor for maintaining trust in academic settings.
• The low latency demonstrates feasibility for real‑time monitoring of large lecture halls or remote proctoring setups.

Practical Implications

• Scalable Proctoring: Universities can deploy the model on existing surveillance infrastructure without needing expensive dedicated hardware.
• Cost Reduction: Automating detection cuts down on human invigilator hours and reduces the need for manual video review.
• Integration Friendly: Because the stages are modular, developers can replace YOLOv8n with a newer detector or swap RexNet‑150 for a transformer‑based classifier as they become available.
• Privacy‑First Workflow: The private‑by‑design reporting mechanism aligns with GDPR‑like regulations, making it easier for institutions to adopt AI‑assisted monitoring without legal pushback.
• Open‑Source Foundations: The authors release code and model weights, enabling rapid prototyping and community‑driven improvements (e.g., adding audio cues or multi‑frame temporal analysis).

Limitations & Future Work

• Single‑frame focus: The current system ignores temporal cues that could catch subtle cheating patterns spread across frames.
• Audio exclusion: No sound analysis; integrating microphone data could detect whispering or device usage.
• Dataset bias: Although diverse, the dataset may still under‑represent certain classroom layouts or cultural contexts, potentially affecting generalization.
• Explainability: The black‑box nature of deep classifiers isn’t addressed; future work could add visual explanations (e.g., Grad‑CAM) to help invigilators understand why a flag was raised.

By tackling these points, the framework could evolve into a comprehensive, multimodal exam integrity platform that balances accuracy, speed, and ethical responsibility.

Authors

• Van-Truong Le
• Le-Khanh Nguyen
• Trong-Doanh Nguyen

Paper Information

• arXiv ID: 2604.16234v1
• Categories: cs.CV, cs.AI
• Published: April 17, 2026
• PDF: Download PDF